ignite-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Yakov Zhdanov <yzhda...@apache.org>
Subject Re: NoopSwapSpaceSpi and dynamically started caches
Date Thu, 18 Feb 2016 10:43:43 GMT
Konstantin,

Here are my comments.

1. I would force same swap configuration on all servers. Therefore, I would
add "checkClient" attribute to annotation which is "true" by default
2. I think all users should properly configure swap along the cluster if
they want to handle possible memory overflow.

--Yakov

2016-02-18 12:35 GMT+03:00 Konstantin Margorin <ruskim@gmail.com>:

> Hi all.
>
> Sorry for asking again, but after some experiments I want to discuss
> several options for grid swap policy. Some of them have odd side effects.
> It's better to explicitly write them down to select the best solution.
>
> 1. Strict policy.
> - @IgniteSpiConsistencyChecked(optional = false) used in FileSwapSpaceSpi
> and NoopSwapSpaceSpi. Nodes with different swap providers could not coexist
> in the grid.
> - Cache with swap space could be start, only if all nodes in the grid has
> FileSwapSpaceSpi provider enabled.
> In this case if part of nodes in the grid use FileSwapSpaceSpi and other
> part use NoopSwapSpaceSpi, than only random part of them will start. If the
> first started node has FileSwapSpaceSpi, then all nodes with
>  NoopSwapSpaceSpi will fail to start, and vice versa.
> Interesting thing is that client nodes ( with Ignition.setClientMode(true))
> should also has the same swap space provider as others nodes in the grid.
> I.e. If nodes in the grid use FileSwapSpaceSpi, client will fail to start
> if it use  NoopSwapSpaceSpi. That is strange, because client node does not
> store any data, so swap policy is irrelevant to them.
> Another interesting thing with client nodes is that client node with
> NoopSwapSpaceSpi will never be able to start cache with swap enabled:
> - if all nodes in the grid use FileSwapSpaceSpi, than client node simply
> will fail to start
> - if all nodes in the grid use NoopSwapSpaceSpi, than cache with swap will
> not start on such nodes
> 2. Soft policy
> - @IgniteSpiConsistencyChecked(optional = true) used in FileSwapSpaceSpi
> and NoopSwapSpaceSpi. Nodes with different swap providers could coexist,
> warning is printed in case of different providers.
> - Cache with swap space could be start, only if all nodes in the grid has
> FileSwapSpaceSpi provider enabled.
> In this case success of dynamic cache with swap creation depends on order
> in which the nodes start. Suppose we have node with swap (nSwap), node
> without swap (nNoSwap) and a client, who just after start creates dynamic
> cache with swap (cSwap).
> If nSwap and nNoSwap started before cSwap, than cSwap will fail to create
> cache, because not all nodes in the grid support swap space.
> If nSwap started first, than started cSwap. cSwap will create cache
> succesfuly. After that nNoSwap will start. And we have configuration now,
> when nodes with different swap policies are coexist in grid, but cache with
> swap successfully started.
> 3. Another option, initially suggested by Artem, is always use
> FileSwapSpaceSpi if user doesn't provide SwapSpaceSpi explicitly. Probably
> will get higher resources usage in this case.
> 4. Leave it as now, but with logging:
> - @IgniteSpiConsistencyChecked(optional = true)
> - log when cache with swap enabled started and there some nodes without
> swap in the grid
> Which option should we choose?
>
>
> On Tue, Feb 16, 2016 at 2:55 PM, Yakov Zhdanov <yzhdanov@apache.org>
> wrote:
>
> > I think, yes
> >
> > --Yakov
> >
> > 2016-02-16 14:19 GMT+03:00 Konstantin Margorin <ruskim@gmail.com>:
> >
> > > Should we use optional=false in @IgniteSpiConsistencyChecked ?
> > >
> > > On Tue, Feb 16, 2016 at 1:15 PM, Dmitriy Setrakyan <
> > dsetrakyan@apache.org>
> > > wrote:
> > >
> > > > Sounds good to me.
> > > >
> > > > On Tue, Feb 16, 2016 at 2:13 AM, Yakov Zhdanov <yzhdanov@apache.org>
> > > > wrote:
> > > >
> > > > > Let's require that all nodes should have same swap space SPI. I
> think
> > > we
> > > > > can put @GridSpiConsistencyChecked on file swap space SPI (any
> other
> > > > > suggestions?).
> > > > >
> > > > > --Yakov
> > > > >
> > > > > 2016-02-15 23:31 GMT+03:00 Konstantin Margorin <ruskim@gmail.com>:
> > > > >
> > > > > > Probably I can handle this ticket, but I want to clarify one
> thing.
> > > > > >
> > > > > > Suppose dynamic cache with enabled swap should be partitioned
> > between
> > > > two
> > > > > > nodes with backups=1. First node started with FileSwapSpaceSpi,
> > > second
> > > > > with
> > > > > > NoopSwapSpaceSpi.
> > > > > >
> > > > > > How cache should start in this case?
> > > > > >
> > > > > >
> > > > > > On Mon, Feb 15, 2016 at 3:45 PM, Artem Shutak <
> > ashutak@gridgain.com>
> > > > > > wrote:
> > > > > >
> > > > > > > I've created a newbie issue for that:
> > > > > > > https://issues.apache.org/jira/browse/IGNITE-2650.
> > > > > > >
> > > > > > > -- Artem --
> > > > > > >
> > > > > > > On Wed, Feb 10, 2016 at 8:53 PM, Dmitriy Setrakyan <
> > > > > > dsetrakyan@apache.org>
> > > > > > > wrote:
> > > > > > >
> > > > > > > > Artem,
> > > > > > > >
> > > > > > > > I like the option with proper exception in case if
a dynamic
> > > cache
> > > > is
> > > > > > > > started with swap enabled while there was no swap
configured
> in
> > > > > Ignite.
> > > > > > > >
> > > > > > > > Starting FileSwapSpaceSpi all the time is not a good
option
> as
> > > user
> > > > > > will
> > > > > > > > face the same problem when trying to change to a different
> SPI.
> > > > > > > >
> > > > > > > > D.
> > > > > > > >
> > > > > > > > On Wed, Feb 10, 2016 at 4:39 AM, Artem Shutak <
> > > > ashutak@gridgain.com>
> > > > > > > > wrote:
> > > > > > > >
> > > > > > > > > Igniters,
> > > > > > > > >
> > > > > > > > > I've noticed an issue with NoopSwapSpaceSpi and
dynamically
> > > > started
> > > > > > > > cashes.
> > > > > > > > >
> > > > > > > > > We have the following logic on node start:
> > > > > > > > >
> > > > > > > > >             if (cfg.getSwapSpaceSpi() == null)
{
> > > > > > > > >                 boolean needSwap = false;
> > > > > > > > >
> > > > > > > > >                 if (cfg.getCacheConfiguration()
!= null &&
> > > > > > > > > !Boolean.TRUE.equals(cfg.isClientMode())) {
> > > > > > > > >                     for (CacheConfiguration c
:
> > > > > > > > > cfg.getCacheConfiguration()) {
> > > > > > > > >                         if (c.isSwapEnabled())
{
> > > > > > > > >                             needSwap = true;
> > > > > > > > >
> > > > > > > > >                             break;
> > > > > > > > >                         }
> > > > > > > > >                     }
> > > > > > > > >                 }
> > > > > > > > >
> > > > > > > > >                 cfg.setSwapSpaceSpi(needSwap
? new
> > > > > > FileSwapSpaceSpi() :
> > > > > > > > new
> > > > > > > > > NoopSwapSpaceSpi());
> > > > > > > > >             }
> > > > > > > > >
> > > > > > > > > According to the code above, if IgniteConfiguration
doesn't
> > > > contain
> > > > > > > cache
> > > > > > > > > configurations or all configured caches configured
without
> > > swap,
> > > > > then
> > > > > > > > node
> > > > > > > > > will be started with NoopSwapSpaceSpi.
> > > > > > > > > As a result in this case, when we start dynamic
cache with
> > > > enabled
> > > > > > swap
> > > > > > > > > then actually swap doesn't work.
> > > > > > > > >
> > > > > > > > > There is simple workaround - user have to explicitly
set
> > > > > > > FileSwapSpaceSpi
> > > > > > > > > to configuration. But now, there is no any warning
or
> > > exception.
> > > > > > > > >
> > > > > > > > > What we can do:
> > > > > > > > >
> > > > > > > > >    1. Throw an exception on dynamic cache start
(with swap
> > > > enabled)
> > > > > > > > >    if NoopSwapSpaceSpi is configured.
> > > > > > > > >    2. Actually, I don't see why we need NoopSwapSpaceSpi
at
> > > all.
> > > > I
> > > > > > > think
> > > > > > > > we
> > > > > > > > >    should configure FileSwapSpaceSpi always if
user doesn't
> > > > provide
> > > > > > any
> > > > > > > > >    SwapSpaceSpi explicitly.
> > > > > > > > >
> > > > > > > > > Thoughts?
> > > > > > > > >
> > > > > > > > > Thanks,
> > > > > > > > > -- Artem --
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > >
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message